Thinking Machines released Inkling on 15 July 2026: a Mixture-of-Experts transformer, 975B total parameters, 41B active, a 1M-token context window, pretrained on 45 trillion tokens of text, images, audio and video. Open weights, with a lighter Inkling-Small preview at 12B active alongside it. As WIRED noted in its write-up, the lab was blunt that this is not the strongest model available — it is meant to be a good base. That distinction is the whole story for an architecture office.

Read the topology, not the leaderboard

Strip the announcement to its dependency graph and three edges matter. First, active parameters: 41B of 975B fire per token, so inference cost tracks the small number while capability tracks the large one. Second, the encoder-free multimodal path — audio arrives as dMel spectrograms (Bai et al., 2024), images as 40×40-pixel patches through a four-layer hMLP (Touvron et al., 2022), both embedded straight into the same token stream as text. There is no bolted-on vision tower to keep in sync. Third, controllable thinking effort: Thinking Machines reports Inkling spending one third the tokens of Nemotron 3 Ultra for equal performance on Terminal Bench 2.1.

That third edge is the one offices underweight. Cost and latency are the binding constraints on any workflow you run ten thousand times, not peak benchmark score. A model you can dial down for clash triage and dial up for a tender clause is a different infrastructure object than a model with one speed.

←TODAY: Open weights at 975B still need a GPU cluster; nobody fine-tunes this on the office workstation.
→3012: The offices that survived owned their model weights the way they once owned their detail library.
Fulcrum: Renting capability is cheap until the vendor deprecates the endpoint your BEP depends on — ownership is a resilience decision, not a cost one.

Where the queue actually builds

The multimodal claim is the part with real AEC surface, and it comes with numbers. On Charxiv RQ — chart and diagram reasoning — Inkling scores 78.1%, rising to 82.0% when it is allowed a Python tool to zoom and crop, against 80.2% for the closed Gemini 3.1 Pro. On MMMU Pro it reaches 73.5%. Charts, diagrams and mathematical visual reasoning are exactly the shape of a Schnitt, a Bauleitung photo, a site voice memo.

But an encoder-free patch pipeline reasoning over a scanned drawing is not the same operation as attention over an IFC element graph — as our concept panel on attention sets out, the operator is identical, the tokens are not. Feed it geometry as geometry and you get structure; feed it a JPEG of geometry and you get plausible description. Know which one you built.

The calibration work is the quieter win. Thinking Machines trained Inkling with reinforcement learning against proper scoring rules and abstention-aware rewards, and reports 61.1 ± 0.79 on the ForecastBench Brier Index without search — level with Gemini 3.1 Pro and above Claude Opus 4.8 at 54.6. A model that says “I don’t know” about a fire-rating detail is worth more on a Bauleitung than one that confabulates fluently.

The honest risk: an open-weights base you fine-tune is a base you now maintain. Weights drift out of date, your eval set rots, and the person who ran the training job leaves. That is a durable dependency, not a free lunch.

Atelier: The Büro question in 2026 is not “which model” but “which of our tasks is stable enough to be worth training on”. Repetitive, well-specified, high-volume work — room-schedule normalisation, LOIN compliance checks, submittal triage — is fine-tune territory. One-off design judgment is not. Monday move: pick your single highest-volume repetitive text task, export 200 real past examples with their correct outputs, and put them in one JSONL file. That corpus, not the model, is the asset — and you can build it before deciding on any vendor.

Hack: Turn your office’s past decisions into a fine-tuning corpus you actually own. Take any table of past inputs and approved outputs and emit the chat-format JSONL that Tinker, HuggingFace TRL, and every other trainer eats:

import json, csv
with open("room_schedule_fixes.csv") as f, open("train.jsonl","w") as out:
    for r in csv.DictReader(f):
        msgs=[{"role":"user","content":r["raw"]},{"role":"assistant","content":r["approved"]}]
        out.write(json.dumps({"messages":msgs})+"\n")

Run it, then read ten random lines back. If the approved column is inconsistent, you have found a documentation problem, not a model problem — and that is worth more than the training run.

The dependency you cannot see yet

I spent thirteen years drawing dependency graphs for systems that were declared resilient right up until the cooling failed. The pattern repeats in software procurement: teams map the architecture diagram, which shows intent, and skip the dependency graph, which shows reality. If a single vendor endpoint sits underneath your model-checking, your specification drafting and your client correspondence, you have one point of failure wearing three costumes. Open weights are one way to cut that edge. So is keeping the corpus.

Because the interesting move here is not downloading a 975B model. It is noticing that Thinking Machines had Inkling write and run its own fine-tuning job on Tinker, and that it sustained 40 iterations of review feedback from GPT Codex on a single artifact. Capability is compounding on the training side, which means the scarce resource shifts from model access to clean, labelled, domain-specific data. Architecture offices sit on decades of exactly that, badly filed.

So file it. Start the JSONL today.

Sources & Further Reading

• Primary: Thinking Machines — Introducing Inkling
• Reinforcing: WIRED — Thinking Machines Lab Drops Its First Model